Sparse Optimization Driven Deep Unfolding Network for DOA Estimation

Direction-of-Arrival (DOA) estimation is widely applied in acoustic source localization. Recent deep learning methods have achieved satisfactory estimation accuracy, attributed to their excellent nonlinear description capabilities in exploiting latent structures embedded in datasets. However, the prior knowledge of array structure is often ignored in the design stage of these networks, resulting in a compromise on model transparency. To address this issue, we propose an interpretable, sparse optimization driven framework for DOA estimation called Array Manifold Integrated LISTA (AMI-LISTA). Firstly, the array manifold matrix is fixed as a constant dictionary in every iterative step and integrated into sparse optimization backbone network LISTA, and thus the intrinsic knowledge of array structure is maintained to ensure network interpretability. Secondly, our approach introduces a layer-fused strategy into the loss function. The layer-fused mechanism facilitates incremental improvements and enhances training stability, leading to faster convergence. Numerical experiments confirm the excellent interpretability of AMI-LISTA and its superior performance, particularly under small angle interval conditions.